An off switch for dual use knowledge in AI models

Anthropic researchers have developed a method to systematically isolate and deactivate specific dangerous knowledge within large language models without degrading their general performance. This technique allows developers to target harmful information, such as instructions for constructing biological weapons, while preserving the model's overall reasoning and language capabilities. By enabling more granular control over model safety, this approach offers a potential solution to the dual-use dilemma where powerful AI systems must balance helpfulness with the prevention of misuse.